In this paper, a Supervised Extreme Learning Machine-based Auto-Encoder (SELM-AE) is proposed for discriminative Feature Learning. Different from traditional ELM-AE (designed based on data information X only), SELM-AE is designed based on both data information X and label information T. In detail, SELM-AE not only minimizes the reconstruction error of input data but also minimizes the intra-class distance and maximizes the inter-class distance in the new feature space. Under this way, the new data representation extracted by proposed SELM-AE is more discriminative than traditional ELM-AE for further classification. Then multiple SELM-AEs are stacked layer by layer to develop a new multi-layer perceptron (MLP) network called ML-SAE-ELM. Benefit from SELM-AE, the proposed ML-SAE-ELM is highly effective on classification than ELM-AE based MLP. Moreover, different from ELM-AE based MLP that requires large number of hidden nodes to achieve satisfactory accuracy, ML-SAE-ELM usually takes very small number of hidden nodes on both feature learning and classification stages to achieve better accuracy, which highly lightens the network memory requirement. The proposed method has been evaluated over 13 benchmark binary and multi-class datasets and one complicated image dataset. As shown in the experimental results, through the visualization of data representation, the proposed SELM-AE extracts more discriminative data representation than ELM-AE. Moreover, the shallow ML-SAE-ELM with smaller hidden nodes achieves higher classification accuracy than hierarchical ELM (a commonly used effective ELM-AE based MLP) on most evaluated datasets.